Uncertainty-driven Ensemble Forecasting of QoS in Software Defined Networks Kostas Kolomvatsos1, Christos Anagnostopoulos2, Angelos Marnerides3, Qiang Ni3, Stathes Hadjiefhtymiades4, Dimitrios Pezaros2 1University of Thessaly, 2University of Glasgow, 3University of Lancaster, 4University of Athens 22nd IEEE Symposium on Computers and Communications 03-06 July 2017, Heraclion, Crete, Greece

Outline SDNs Management The proposed framework The ensemble forecasting scheme The uncertainty management mechanism Experimental evaluation

SDNs Management (1/2) SDN controllers are responsible for performing various network tasks They are connected to applications through the northbound interface and to devices through the southbound interface They incorporate monitoring functionalities to collect time-series network performance data Example metrics: latency, link utilization, switch buffer occupancy, etc

SDNs Management (2/2) During the functioning the amount of data becomes huge We want to derive predictive analytics on top of the data Our aim is to use the analytics to secure the QoS

The proposed framework We manage various Network Performance Parameters (NPPs) We propose a module that supports real time decision making Our module involves: A Type-2 Fuzzy Logic System (T2FLS) for uncertainty management (it derives alerts about the presence of QoS violation events) The T2FLS derives the Potential of Violation (PoV) A combination of responses derived through large-scale predictive analytics A combination of multiple aggregated time series forecasting results Our goal: Provide forecasting analytics to the SDN controller For each NPP we provide an aggregated value The T2FLS informs the SDN controller for the presence of an event

The ensemble forecasting scheme (1/3) We consider a set of estimators Examples: auto-regressive estimator, double and triple exponential smoothing, weighted and cumulative moving average, etc We adopt 28 estimators The ensemble scheme involves the aggregation of multiple estimators for each NPP The final aggregated value is derived through an aggregation function on top of historical values

The ensemble forecasting scheme (2/3) We adopt a linear aggregation function Each estimator has a specific weight Our model produces a vector of aggregated estimations (a value for each NPP) Weights are defined based on estimators’ performance The performance is affected by the estimation error is the estimation and is the real observation for the k-th NPP

The ensemble forecasting scheme (3/3) We adopt a sliding window approach We determine the weight of an estimator based on its performance in the window Weights are based on the average forecasting error We define a convex combination rule for weights definition The mechanism assigns high weight to estimators with low average error

Uncertainty management (1/2) We want to ‘fire’ the update on the orchestration process of the SDN controller Uncertainty is present on how the aggregated estimation depicts a high potential of QoS violation We propose the T2FLS for such purposes The T2FLS linearly maps the inputs to the outputs It adopts as et of rules are the inputs, are the outputs and and are membership functions

Uncertainty management (2/2) Membership functions in Type-2 FLSs are intervals Our T2FLS has |M| inputs; M is the set of NPPs Each input corresponds to the aggregated measurement for an NPP We consider three linguistic values: Low, Medium, High The output is the PoV When PoV is over a pre-defined threshold the T2FLS fires an event to the SDN controller

Experimental evaluation (1/3) Experimental setup We focus on: link utilization β, average switches buffer size ε and average latency α We adopt two distributions to produce values for each metric: Uniform, Exponential (λ=0.5, λ=2.0) We adopt three decision thresholds: βΤ, εΤ, αΤ When NPPs violate the thresholds, there is an indication of QoS violation We adopt known performance metrics for our model like: - Precision - Recall - Accuracy - F-measure TP: true-positive, TN: true-negative, FP: false-positive, FN: false-negative

Experimental evaluation (2/3) We run 1,000 simulations for 1,000 rounds per simulation At t, we collect the realization for NPPs and execute the proposed scheme Three scenarios are evaluated: Scenario A: βΤ = εΤ = αΤ = 0.5 Scenario B: βΤ = 0.3, εΤ = αΤ = 0.7 Scenario C: βΤ = 0.7, εΤ = αΤ = 0.3

Experimental evaluation (3/3) Results ( is the number of estimators)

Thank you for your attention!